[OldRocketGuy]

A 128 bit vector engine was mentioned in the keynote on Monday. It was said to be a replacement for the AltiVec on the old architecture. Is this vector engine part of Intel's Woodcrest CPU or support chip? Or is it Apple proprietary silicon added to the quad architecture? I have not seen any discussion about this feature.

[Tomchu]

That would simply be SSE/SSE2/SSE3.

[OldRocketGuy]

Thanks, I found the Intel architecture document that explained it.

[Simon]

Speaking of SSE3, do we have reason to believe Apple has gone through their code (OS, iApps, pro apps) and made sure the Intel code is optimized for SSE3 just as the PPC code is for Altivec? Or are we not yet seeing the full SSE3 benefits because the code was optimized for Altivec in the past?

[OreoCookie]

SSE seems to be harder to program than AltiVec -- which was one of the reasons why it was so successful in speeding up certain operations, like encoding videos or mp3s. The SSE units are also used for fp operations by default.

[Tomchu]

A good enough compiler should be able to vectorize to a decent extent whatever code is vectorizable. I'm not sure what kind of work Apple/Intel have done on GCC 4 in this regard.

[OreoCookie]

The Gnu compiler suite does autovectorization for both, PPC and x86 cpus. Other compilers might do a better job, though. However, to really get optimal performance, you have to tweak the code by hand. This is particularly beneficial for audio and video encoding, complex scientific calculations and pretty much anything where a relatively small part of the code uses up a lot of computation time.

[Don Pickett]

I was under the impression that one of the reasons Apple was introducing the Core frameworks (CoreAudio, Video, etc.) was to allow performance which equalled that of AltiVec-enhanced code without needing the AltiVec code. Fr'example, CoreVideo dumps most of its processing to the video card, etc.

[mduell]

I would hope Apple is using ICC for the OS and high performance apps... or at least flogging/helping Intel to port it if it hasn't been ported yet.

[milhous]

would there be any incentives/advantages for apple to just use the intel compiler over GCC?

[Don Pickett]

Originally Posted by milhous

would there be any incentives/advantages for apple to just use the intel compiler over GCC?

I believe that, because it's Intel's compiler it optimized for their chips very well.

[mduell]

Originally Posted by milhous

would there be any incentives/advantages for apple to just use the intel compiler over GCC?

Yes. Plenty of benchmarks on google to demonstrate that.

[OldRocketGuy]

There are some key math libraries that are used by many scientific, engineering and statistical applications that would probably greatly improve performance if they were compiled with tweaks and available in Xcode. I'm thinking about stuff like the BLAS, LAPACK, FFTW and such. It would make the Mac Pro an even bigger winner than it already is.

It would behoove Apple to make this available. I'm sure that shops such as Los Alamos and Livermore will do this sort of work for their own benefit. Given that they are government shops, such code should be Open Source and available to all.

[OreoCookie]

You can always compile them yourself. The main reason Apple uses GCC and not some other compiler is to keep cross-platform compatibility. Apple didn't just keep a copy of OS X running on x86 hardware to switch to x86 cpus, but to keep cross-platform compatibility -- so they could have switched to any platform.

Since Apple needs to support PowerPCs for at least five or six more years from now, they need to stick to gcc for now. Also, who knows what will happen in 5 or 10 years from now? Being able to port the OS to another cpu architecture is a good thing. Everybody who needs to optimize certain libraries, they are available either as source code or highly optimized (for money).